Multi-scale residual neural network for image steganalysis

Deep steganalysis networks (DSNs) has made a great progress in detection performance. However, most of the deep image steganalysis networks are not complete end-to-end models. In their approach, traditional hand-crafted features are employed to pre-process the images, which can obtain the high-frequency noise residuals to alleviate the interference of the image content. To avoid relying on domain knowledge of deep learning-based methods, we design an end-to-end deep image steganalysis neural model that combines multi-scale feature extraction and residual fusion modules. Firstly, we use the standard convolution kernels of different sizes to extract the features of different scale receptive fields in the input image. Then, depth-wise convolution is used to independently models the inner-channel correlations of multi-scale features and retains the discriminative statistical characteristics of each channel. The residual fusion technique is introduced to aggregate hierarchical feature and strengthen information representation in the network. Experimental results show that, compared with the existing classical deep image steganalysis networks, the proposed steganalysis scheme has a great improvement in steganalysis error rates.